Fig. 2.

Illustrating probabilistic concepts of our approach: (a) three Gaussian components with different means and equal covariances used to generate the points that are displayed over contours of their marginal distribution. The points in each desired (gold-standard) cluster are shown in a different color and shapes; (b,c) the resulting groups after doing an unconstrained spectral clustering lying on contours of their corresponding posterior distribution; (d) the resulting groups shown together over contours of the posterior entropy h, with the encircled point as the one with the largest h; (e) the resulting groups over contours of the density weighted entropy ϕ, with the encircled point as the one with maximum ϕ (i.e. q₀); (f) the resulting groups over contours of the inverse-density weighted entropy ψ, with the encircled points having the minimum ψ and shortest distance to q₀ in each cluster (i.e. q₁ and q₂).